Motivated by methodological issues encountered in the analysis of the Burden of Obstructive Lung Disease (BOLD) study, I used a Bayesian approach to answer two important questions on the impact of multiple risk factors on chronic airflow obstruction (CAO) in different regions of the world.
First, I developed a model to improve the precision of the estimates of two measures of population impact of a risk factor on disease burden: Population Attributable Fraction (PAF) and Population Attributable Risk (PAR). These measures need to be estimated locally, as they depend on both relative risk (RR) and distribution of the risk factor in a given population, which is challenging in BOLD sites with limited data. I developed a Bayesian hierarchical model that borrows information on mean and variance of the RR across sites, leading to more precise estimates of site-specific RRs and therefore of local PAFs/PARs. I used this method to investigate the local impact of several risk factors on CAO burden. Finally, I implemented it in a user-friendly tool soon freely available online for others to use, including researchers with no expertise in Bayesian methods.
Secondly, I collaborated with colleagues to develop a Bayesian method to impute data that are systematically missing in some datasets when combining information across multiple datasets. I applied this method to estimate the association of poverty with CAO using data from all the 41 BOLD sites, through imputation of data on poverty in 20 sites that had not collected this information. We implemented this method in a user-friendly online tool, BIMAM.
This thesis contributes to the field of respiratory epidemiology by demonstrating the different role played by modifiable risk factors for CAO in different parts of the world, and by contributing to the methodological improvement of the analysis of multi-centre international studies like BOLD.